1 //===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 /// This file implements a TargetTransformInfo analysis pass specific to the
11 /// X86 target machine. It uses the target's detailed information to provide
12 /// more precise answers to certain TTI queries, while letting the target
13 /// independent and default TTI implementations handle the rest.
15 //===----------------------------------------------------------------------===//
18 #include "X86TargetMachine.h"
19 #include "llvm/Analysis/TargetTransformInfo.h"
20 #include "llvm/IR/IntrinsicInst.h"
21 #include "llvm/Support/Debug.h"
22 #include "llvm/Target/CostTable.h"
23 #include "llvm/Target/TargetLowering.h"
26 #define DEBUG_TYPE "x86tti"
28 // Declare the pass initialization routine locally as target-specific passes
29 // don't have a target-wide initialization entry point, and so we rely on the
30 // pass constructor initialization.
32 void initializeX86TTIPass(PassRegistry &);
37 class X86TTI final : public ImmutablePass, public TargetTransformInfo {
38 const X86Subtarget *ST;
39 const X86TargetLowering *TLI;
41 /// Estimate the overhead of scalarizing an instruction. Insert and Extract
42 /// are set if the result needs to be inserted and/or extracted from vectors.
43 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
46 X86TTI() : ImmutablePass(ID), ST(nullptr), TLI(nullptr) {
47 llvm_unreachable("This pass cannot be directly constructed");
50 X86TTI(const X86TargetMachine *TM)
51 : ImmutablePass(ID), ST(TM->getSubtargetImpl()),
52 TLI(TM->getTargetLowering()) {
53 initializeX86TTIPass(*PassRegistry::getPassRegistry());
56 void initializePass() override {
60 void getAnalysisUsage(AnalysisUsage &AU) const override {
61 TargetTransformInfo::getAnalysisUsage(AU);
64 /// Pass identification.
67 /// Provide necessary pointer adjustments for the two base classes.
68 void *getAdjustedAnalysisPointer(const void *ID) override {
69 if (ID == &TargetTransformInfo::ID)
70 return (TargetTransformInfo*)this;
74 /// \name Scalar TTI Implementations
76 PopcntSupportKind getPopcntSupport(unsigned TyWidth) const override;
80 /// \name Vector TTI Implementations
83 unsigned getNumberOfRegisters(bool Vector) const override;
84 unsigned getRegisterBitWidth(bool Vector) const override;
85 unsigned getMaximumUnrollFactor() const override;
86 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind,
87 OperandValueKind) const override;
88 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp,
89 int Index, Type *SubTp) const override;
90 unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
91 Type *Src) const override;
92 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
93 Type *CondTy) const override;
94 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
95 unsigned Index) const override;
96 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
97 unsigned AddressSpace) const override;
99 unsigned getAddressComputationCost(Type *PtrTy,
100 bool IsComplex) const override;
102 unsigned getReductionCost(unsigned Opcode, Type *Ty,
103 bool IsPairwiseForm) const override;
105 unsigned getIntImmCost(int64_t) const;
107 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const override;
109 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
110 Type *Ty) const override;
111 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
112 Type *Ty) const override;
117 } // end anonymous namespace
119 INITIALIZE_AG_PASS(X86TTI, TargetTransformInfo, "x86tti",
120 "X86 Target Transform Info", true, true, false)
124 llvm::createX86TargetTransformInfoPass(const X86TargetMachine *TM) {
125 return new X86TTI(TM);
129 //===----------------------------------------------------------------------===//
133 //===----------------------------------------------------------------------===//
135 X86TTI::PopcntSupportKind X86TTI::getPopcntSupport(unsigned TyWidth) const {
136 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
137 // TODO: Currently the __builtin_popcount() implementation using SSE3
138 // instructions is inefficient. Once the problem is fixed, we should
139 // call ST->hasSSE3() instead of ST->hasPOPCNT().
140 return ST->hasPOPCNT() ? PSK_FastHardware : PSK_Software;
143 unsigned X86TTI::getNumberOfRegisters(bool Vector) const {
144 if (Vector && !ST->hasSSE1())
152 unsigned X86TTI::getRegisterBitWidth(bool Vector) const {
154 if (ST->hasAVX()) return 256;
155 if (ST->hasSSE1()) return 128;
165 unsigned X86TTI::getMaximumUnrollFactor() const {
169 // Sandybridge and Haswell have multiple execution ports and pipelined
177 unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
178 OperandValueKind Op1Info,
179 OperandValueKind Op2Info) const {
180 // Legalize the type.
181 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
183 int ISD = TLI->InstructionOpcodeToISD(Opcode);
184 assert(ISD && "Invalid opcode");
186 static const CostTblEntry<MVT::SimpleValueType>
187 AVX2UniformConstCostTable[] = {
188 { ISD::SDIV, MVT::v16i16, 6 }, // vpmulhw sequence
189 { ISD::UDIV, MVT::v16i16, 6 }, // vpmulhuw sequence
190 { ISD::SDIV, MVT::v8i32, 15 }, // vpmuldq sequence
191 { ISD::UDIV, MVT::v8i32, 15 }, // vpmuludq sequence
194 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
196 int Idx = CostTableLookup(AVX2UniformConstCostTable, ISD, LT.second);
198 return LT.first * AVX2UniformConstCostTable[Idx].Cost;
201 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTable[] = {
202 // Shifts on v4i64/v8i32 on AVX2 is legal even though we declare to
203 // customize them to detect the cases where shift amount is a scalar one.
204 { ISD::SHL, MVT::v4i32, 1 },
205 { ISD::SRL, MVT::v4i32, 1 },
206 { ISD::SRA, MVT::v4i32, 1 },
207 { ISD::SHL, MVT::v8i32, 1 },
208 { ISD::SRL, MVT::v8i32, 1 },
209 { ISD::SRA, MVT::v8i32, 1 },
210 { ISD::SHL, MVT::v2i64, 1 },
211 { ISD::SRL, MVT::v2i64, 1 },
212 { ISD::SHL, MVT::v4i64, 1 },
213 { ISD::SRL, MVT::v4i64, 1 },
215 { ISD::SHL, MVT::v32i8, 42 }, // cmpeqb sequence.
216 { ISD::SHL, MVT::v16i16, 16*10 }, // Scalarized.
218 { ISD::SRL, MVT::v32i8, 32*10 }, // Scalarized.
219 { ISD::SRL, MVT::v16i16, 8*10 }, // Scalarized.
221 { ISD::SRA, MVT::v32i8, 32*10 }, // Scalarized.
222 { ISD::SRA, MVT::v16i16, 16*10 }, // Scalarized.
223 { ISD::SRA, MVT::v4i64, 4*10 }, // Scalarized.
225 // Vectorizing division is a bad idea. See the SSE2 table for more comments.
226 { ISD::SDIV, MVT::v32i8, 32*20 },
227 { ISD::SDIV, MVT::v16i16, 16*20 },
228 { ISD::SDIV, MVT::v8i32, 8*20 },
229 { ISD::SDIV, MVT::v4i64, 4*20 },
230 { ISD::UDIV, MVT::v32i8, 32*20 },
231 { ISD::UDIV, MVT::v16i16, 16*20 },
232 { ISD::UDIV, MVT::v8i32, 8*20 },
233 { ISD::UDIV, MVT::v4i64, 4*20 },
236 // Look for AVX2 lowering tricks.
238 if (ISD == ISD::SHL && LT.second == MVT::v16i16 &&
239 (Op2Info == TargetTransformInfo::OK_UniformConstantValue ||
240 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue))
241 // On AVX2, a packed v16i16 shift left by a constant build_vector
242 // is lowered into a vector multiply (vpmullw).
245 int Idx = CostTableLookup(AVX2CostTable, ISD, LT.second);
247 return LT.first * AVX2CostTable[Idx].Cost;
250 static const CostTblEntry<MVT::SimpleValueType>
251 SSE2UniformConstCostTable[] = {
252 // We don't correctly identify costs of casts because they are marked as
254 // Constant splats are cheaper for the following instructions.
255 { ISD::SHL, MVT::v16i8, 1 }, // psllw.
256 { ISD::SHL, MVT::v8i16, 1 }, // psllw.
257 { ISD::SHL, MVT::v4i32, 1 }, // pslld
258 { ISD::SHL, MVT::v2i64, 1 }, // psllq.
260 { ISD::SRL, MVT::v16i8, 1 }, // psrlw.
261 { ISD::SRL, MVT::v8i16, 1 }, // psrlw.
262 { ISD::SRL, MVT::v4i32, 1 }, // psrld.
263 { ISD::SRL, MVT::v2i64, 1 }, // psrlq.
265 { ISD::SRA, MVT::v16i8, 4 }, // psrlw, pand, pxor, psubb.
266 { ISD::SRA, MVT::v8i16, 1 }, // psraw.
267 { ISD::SRA, MVT::v4i32, 1 }, // psrad.
269 { ISD::SDIV, MVT::v8i16, 6 }, // pmulhw sequence
270 { ISD::UDIV, MVT::v8i16, 6 }, // pmulhuw sequence
271 { ISD::SDIV, MVT::v4i32, 19 }, // pmuludq sequence
272 { ISD::UDIV, MVT::v4i32, 15 }, // pmuludq sequence
275 if (Op2Info == TargetTransformInfo::OK_UniformConstantValue &&
278 if (ISD == ISD::SDIV && LT.second == MVT::v4i32 && ST->hasSSE41())
279 return LT.first * 15;
281 int Idx = CostTableLookup(SSE2UniformConstCostTable, ISD, LT.second);
283 return LT.first * SSE2UniformConstCostTable[Idx].Cost;
286 if (ISD == ISD::SHL &&
287 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue) {
289 if ((VT == MVT::v8i16 && ST->hasSSE2()) ||
290 (VT == MVT::v4i32 && ST->hasSSE41()))
291 // Vector shift left by non uniform constant can be lowered
292 // into vector multiply (pmullw/pmulld).
294 if (VT == MVT::v4i32 && ST->hasSSE2())
295 // A vector shift left by non uniform constant is converted
296 // into a vector multiply; the new multiply is eventually
297 // lowered into a sequence of shuffles and 2 x pmuludq.
301 static const CostTblEntry<MVT::SimpleValueType> SSE2CostTable[] = {
302 // We don't correctly identify costs of casts because they are marked as
304 // For some cases, where the shift amount is a scalar we would be able
305 // to generate better code. Unfortunately, when this is the case the value
306 // (the splat) will get hoisted out of the loop, thereby making it invisible
307 // to ISel. The cost model must return worst case assumptions because it is
308 // used for vectorization and we don't want to make vectorized code worse
310 { ISD::SHL, MVT::v16i8, 30 }, // cmpeqb sequence.
311 { ISD::SHL, MVT::v8i16, 8*10 }, // Scalarized.
312 { ISD::SHL, MVT::v4i32, 2*5 }, // We optimized this using mul.
313 { ISD::SHL, MVT::v2i64, 2*10 }, // Scalarized.
314 { ISD::SHL, MVT::v4i64, 4*10 }, // Scalarized.
316 { ISD::SRL, MVT::v16i8, 16*10 }, // Scalarized.
317 { ISD::SRL, MVT::v8i16, 8*10 }, // Scalarized.
318 { ISD::SRL, MVT::v4i32, 4*10 }, // Scalarized.
319 { ISD::SRL, MVT::v2i64, 2*10 }, // Scalarized.
321 { ISD::SRA, MVT::v16i8, 16*10 }, // Scalarized.
322 { ISD::SRA, MVT::v8i16, 8*10 }, // Scalarized.
323 { ISD::SRA, MVT::v4i32, 4*10 }, // Scalarized.
324 { ISD::SRA, MVT::v2i64, 2*10 }, // Scalarized.
326 // It is not a good idea to vectorize division. We have to scalarize it and
327 // in the process we will often end up having to spilling regular
328 // registers. The overhead of division is going to dominate most kernels
329 // anyways so try hard to prevent vectorization of division - it is
330 // generally a bad idea. Assume somewhat arbitrarily that we have to be able
331 // to hide "20 cycles" for each lane.
332 { ISD::SDIV, MVT::v16i8, 16*20 },
333 { ISD::SDIV, MVT::v8i16, 8*20 },
334 { ISD::SDIV, MVT::v4i32, 4*20 },
335 { ISD::SDIV, MVT::v2i64, 2*20 },
336 { ISD::UDIV, MVT::v16i8, 16*20 },
337 { ISD::UDIV, MVT::v8i16, 8*20 },
338 { ISD::UDIV, MVT::v4i32, 4*20 },
339 { ISD::UDIV, MVT::v2i64, 2*20 },
343 int Idx = CostTableLookup(SSE2CostTable, ISD, LT.second);
345 return LT.first * SSE2CostTable[Idx].Cost;
348 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTable[] = {
349 // We don't have to scalarize unsupported ops. We can issue two half-sized
350 // operations and we only need to extract the upper YMM half.
351 // Two ops + 1 extract + 1 insert = 4.
352 { ISD::MUL, MVT::v16i16, 4 },
353 { ISD::MUL, MVT::v8i32, 4 },
354 { ISD::SUB, MVT::v8i32, 4 },
355 { ISD::ADD, MVT::v8i32, 4 },
356 { ISD::SUB, MVT::v4i64, 4 },
357 { ISD::ADD, MVT::v4i64, 4 },
358 // A v4i64 multiply is custom lowered as two split v2i64 vectors that then
359 // are lowered as a series of long multiplies(3), shifts(4) and adds(2)
360 // Because we believe v4i64 to be a legal type, we must also include the
361 // split factor of two in the cost table. Therefore, the cost here is 18
363 { ISD::MUL, MVT::v4i64, 18 },
366 // Look for AVX1 lowering tricks.
367 if (ST->hasAVX() && !ST->hasAVX2()) {
370 // v16i16 and v8i32 shifts by non-uniform constants are lowered into a
371 // sequence of extract + two vector multiply + insert.
372 if (ISD == ISD::SHL && (VT == MVT::v8i32 || VT == MVT::v16i16) &&
373 Op2Info == TargetTransformInfo::OK_NonUniformConstantValue)
376 int Idx = CostTableLookup(AVX1CostTable, ISD, VT);
378 return LT.first * AVX1CostTable[Idx].Cost;
381 // Custom lowering of vectors.
382 static const CostTblEntry<MVT::SimpleValueType> CustomLowered[] = {
383 // A v2i64/v4i64 and multiply is custom lowered as a series of long
384 // multiplies(3), shifts(4) and adds(2).
385 { ISD::MUL, MVT::v2i64, 9 },
386 { ISD::MUL, MVT::v4i64, 9 },
388 int Idx = CostTableLookup(CustomLowered, ISD, LT.second);
390 return LT.first * CustomLowered[Idx].Cost;
392 // Special lowering of v4i32 mul on sse2, sse3: Lower v4i32 mul as 2x shuffle,
393 // 2x pmuludq, 2x shuffle.
394 if (ISD == ISD::MUL && LT.second == MVT::v4i32 && ST->hasSSE2() &&
398 // Fallback to the default implementation.
399 return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty, Op1Info,
403 unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
405 // We only estimate the cost of reverse shuffles.
406 if (Kind != SK_Reverse)
407 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp);
409 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp);
411 if (LT.second.getSizeInBits() > 128)
412 Cost = 3; // Extract + insert + copy.
414 // Multiple by the number of parts.
415 return Cost * LT.first;
418 unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const {
419 int ISD = TLI->InstructionOpcodeToISD(Opcode);
420 assert(ISD && "Invalid opcode");
422 std::pair<unsigned, MVT> LTSrc = TLI->getTypeLegalizationCost(Src);
423 std::pair<unsigned, MVT> LTDest = TLI->getTypeLegalizationCost(Dst);
425 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
427 // These are somewhat magic numbers justified by looking at the output of
428 // Intel's IACA, running some kernels and making sure when we take
429 // legalization into account the throughput will be overestimated.
430 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
431 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
432 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
433 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
434 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
435 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v4i32, 4*10 },
436 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v8i16, 8*10 },
437 { ISD::SINT_TO_FP, MVT::v2f64, MVT::v16i8, 16*10 },
438 // There are faster sequences for float conversions.
439 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
440 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
441 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
442 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
443 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v2i64, 15 },
444 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 15 },
445 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v8i16, 15 },
446 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v16i8, 8 },
449 if (ST->hasSSE2() && !ST->hasAVX()) {
451 ConvertCostTableLookup(SSE2ConvTbl, ISD, LTDest.second, LTSrc.second);
453 return LTSrc.first * SSE2ConvTbl[Idx].Cost;
456 EVT SrcTy = TLI->getValueType(Src);
457 EVT DstTy = TLI->getValueType(Dst);
459 // The function getSimpleVT only handles simple value types.
460 if (!SrcTy.isSimple() || !DstTy.isSimple())
461 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
463 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
464 AVX2ConversionTbl[] = {
465 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
466 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 1 },
467 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
468 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 3 },
469 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
470 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3 },
471 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
472 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 },
473 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
474 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 3 },
475 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
476 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 3 },
477 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
478 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
479 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
480 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 },
482 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 2 },
483 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2 },
484 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 2 },
485 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2 },
486 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 2 },
487 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 4 },
490 static const TypeConversionCostTblEntry<MVT::SimpleValueType>
491 AVXConversionTbl[] = {
492 { ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
493 { ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 4 },
494 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 7 },
495 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 4 },
496 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 7 },
497 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 4 },
498 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
499 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 4 },
500 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i1, 6 },
501 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i1, 4 },
502 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i8, 6 },
503 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i8, 4 },
504 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 6 },
505 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3 },
506 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
507 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 4 },
509 { ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 4 },
510 { ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 4 },
511 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 4 },
512 { ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 4 },
513 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 5 },
514 { ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 4 },
515 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 9 },
517 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i1, 8 },
518 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 8 },
519 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
520 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i32, 1 },
521 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i1, 3 },
522 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 3 },
523 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 3 },
524 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1 },
525 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i1, 3 },
526 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i8, 3 },
527 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i16, 3 },
528 { ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 1 },
530 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i1, 6 },
531 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 5 },
532 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 5 },
533 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i32, 9 },
534 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i1, 7 },
535 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 2 },
536 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2 },
537 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 6 },
538 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i1, 7 },
539 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i8, 2 },
540 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i16, 2 },
541 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 6 },
542 // The generic code to compute the scalar overhead is currently broken.
543 // Workaround this limitation by estimating the scalarization overhead
544 // here. We have roughly 10 instructions per scalar element.
545 // Multiply that by the vector width.
546 // FIXME: remove that when PR19268 is fixed.
547 { ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 2*10 },
548 { ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i64, 4*10 },
550 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 7 },
551 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 },
552 // This node is expanded into scalarized operations but BasicTTI is overly
553 // optimistic estimating its cost. It computes 3 per element (one
554 // vector-extract, one scalar conversion and one vector-insert). The
555 // problem is that the inserts form a read-modify-write chain so latency
556 // should be factored in too. Inflating the cost per element by 1.
557 { ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f32, 8*4 },
558 { ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f64, 4*4 },
562 int Idx = ConvertCostTableLookup(AVX2ConversionTbl, ISD,
563 DstTy.getSimpleVT(), SrcTy.getSimpleVT());
565 return AVX2ConversionTbl[Idx].Cost;
569 int Idx = ConvertCostTableLookup(AVXConversionTbl, ISD, DstTy.getSimpleVT(),
570 SrcTy.getSimpleVT());
572 return AVXConversionTbl[Idx].Cost;
575 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
578 unsigned X86TTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
579 Type *CondTy) const {
580 // Legalize the type.
581 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
585 int ISD = TLI->InstructionOpcodeToISD(Opcode);
586 assert(ISD && "Invalid opcode");
588 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTbl[] = {
589 { ISD::SETCC, MVT::v2f64, 1 },
590 { ISD::SETCC, MVT::v4f32, 1 },
591 { ISD::SETCC, MVT::v2i64, 1 },
592 { ISD::SETCC, MVT::v4i32, 1 },
593 { ISD::SETCC, MVT::v8i16, 1 },
594 { ISD::SETCC, MVT::v16i8, 1 },
597 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTbl[] = {
598 { ISD::SETCC, MVT::v4f64, 1 },
599 { ISD::SETCC, MVT::v8f32, 1 },
600 // AVX1 does not support 8-wide integer compare.
601 { ISD::SETCC, MVT::v4i64, 4 },
602 { ISD::SETCC, MVT::v8i32, 4 },
603 { ISD::SETCC, MVT::v16i16, 4 },
604 { ISD::SETCC, MVT::v32i8, 4 },
607 static const CostTblEntry<MVT::SimpleValueType> AVX2CostTbl[] = {
608 { ISD::SETCC, MVT::v4i64, 1 },
609 { ISD::SETCC, MVT::v8i32, 1 },
610 { ISD::SETCC, MVT::v16i16, 1 },
611 { ISD::SETCC, MVT::v32i8, 1 },
615 int Idx = CostTableLookup(AVX2CostTbl, ISD, MTy);
617 return LT.first * AVX2CostTbl[Idx].Cost;
621 int Idx = CostTableLookup(AVX1CostTbl, ISD, MTy);
623 return LT.first * AVX1CostTbl[Idx].Cost;
626 if (ST->hasSSE42()) {
627 int Idx = CostTableLookup(SSE42CostTbl, ISD, MTy);
629 return LT.first * SSE42CostTbl[Idx].Cost;
632 return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy);
635 unsigned X86TTI::getVectorInstrCost(unsigned Opcode, Type *Val,
636 unsigned Index) const {
637 assert(Val->isVectorTy() && "This must be a vector type");
640 // Legalize the type.
641 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Val);
643 // This type is legalized to a scalar type.
644 if (!LT.second.isVector())
647 // The type may be split. Normalize the index to the new type.
648 unsigned Width = LT.second.getVectorNumElements();
649 Index = Index % Width;
651 // Floating point scalars are already located in index #0.
652 if (Val->getScalarType()->isFloatingPointTy() && Index == 0)
656 return TargetTransformInfo::getVectorInstrCost(Opcode, Val, Index);
659 unsigned X86TTI::getScalarizationOverhead(Type *Ty, bool Insert,
660 bool Extract) const {
661 assert (Ty->isVectorTy() && "Can only scalarize vectors");
664 for (int i = 0, e = Ty->getVectorNumElements(); i < e; ++i) {
666 Cost += TopTTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
668 Cost += TopTTI->getVectorInstrCost(Instruction::ExtractElement, Ty, i);
674 unsigned X86TTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
675 unsigned AddressSpace) const {
676 // Handle non-power-of-two vectors such as <3 x float>
677 if (VectorType *VTy = dyn_cast<VectorType>(Src)) {
678 unsigned NumElem = VTy->getVectorNumElements();
680 // Handle a few common cases:
682 if (NumElem == 3 && VTy->getScalarSizeInBits() == 32)
683 // Cost = 64 bit store + extract + 32 bit store.
687 if (NumElem == 3 && VTy->getScalarSizeInBits() == 64)
688 // Cost = 128 bit store + unpack + 64 bit store.
691 // Assume that all other non-power-of-two numbers are scalarized.
692 if (!isPowerOf2_32(NumElem)) {
693 unsigned Cost = TargetTransformInfo::getMemoryOpCost(Opcode,
694 VTy->getScalarType(),
697 unsigned SplitCost = getScalarizationOverhead(Src,
698 Opcode == Instruction::Load,
699 Opcode==Instruction::Store);
700 return NumElem * Cost + SplitCost;
704 // Legalize the type.
705 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
706 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
709 // Each load/store unit costs 1.
710 unsigned Cost = LT.first * 1;
712 // On Sandybridge 256bit load/stores are double pumped
713 // (but not on Haswell).
714 if (LT.second.getSizeInBits() > 128 && !ST->hasAVX2())
720 unsigned X86TTI::getAddressComputationCost(Type *Ty, bool IsComplex) const {
721 // Address computations in vectorized code with non-consecutive addresses will
722 // likely result in more instructions compared to scalar code where the
723 // computation can more often be merged into the index mode. The resulting
724 // extra micro-ops can significantly decrease throughput.
725 unsigned NumVectorInstToHideOverhead = 10;
727 if (Ty->isVectorTy() && IsComplex)
728 return NumVectorInstToHideOverhead;
730 return TargetTransformInfo::getAddressComputationCost(Ty, IsComplex);
733 unsigned X86TTI::getReductionCost(unsigned Opcode, Type *ValTy,
734 bool IsPairwise) const {
736 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
740 int ISD = TLI->InstructionOpcodeToISD(Opcode);
741 assert(ISD && "Invalid opcode");
743 // We use the Intel Architecture Code Analyzer(IACA) to measure the throughput
744 // and make it as the cost.
746 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblPairWise[] = {
747 { ISD::FADD, MVT::v2f64, 2 },
748 { ISD::FADD, MVT::v4f32, 4 },
749 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
750 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
751 { ISD::ADD, MVT::v8i16, 5 },
754 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblPairWise[] = {
755 { ISD::FADD, MVT::v4f32, 4 },
756 { ISD::FADD, MVT::v4f64, 5 },
757 { ISD::FADD, MVT::v8f32, 7 },
758 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
759 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.5".
760 { ISD::ADD, MVT::v4i64, 5 }, // The data reported by the IACA tool is "4.8".
761 { ISD::ADD, MVT::v8i16, 5 },
762 { ISD::ADD, MVT::v8i32, 5 },
765 static const CostTblEntry<MVT::SimpleValueType> SSE42CostTblNoPairWise[] = {
766 { ISD::FADD, MVT::v2f64, 2 },
767 { ISD::FADD, MVT::v4f32, 4 },
768 { ISD::ADD, MVT::v2i64, 2 }, // The data reported by the IACA tool is "1.6".
769 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "3.3".
770 { ISD::ADD, MVT::v8i16, 4 }, // The data reported by the IACA tool is "4.3".
773 static const CostTblEntry<MVT::SimpleValueType> AVX1CostTblNoPairWise[] = {
774 { ISD::FADD, MVT::v4f32, 3 },
775 { ISD::FADD, MVT::v4f64, 3 },
776 { ISD::FADD, MVT::v8f32, 4 },
777 { ISD::ADD, MVT::v2i64, 1 }, // The data reported by the IACA tool is "1.5".
778 { ISD::ADD, MVT::v4i32, 3 }, // The data reported by the IACA tool is "2.8".
779 { ISD::ADD, MVT::v4i64, 3 },
780 { ISD::ADD, MVT::v8i16, 4 },
781 { ISD::ADD, MVT::v8i32, 5 },
786 int Idx = CostTableLookup(AVX1CostTblPairWise, ISD, MTy);
788 return LT.first * AVX1CostTblPairWise[Idx].Cost;
791 if (ST->hasSSE42()) {
792 int Idx = CostTableLookup(SSE42CostTblPairWise, ISD, MTy);
794 return LT.first * SSE42CostTblPairWise[Idx].Cost;
798 int Idx = CostTableLookup(AVX1CostTblNoPairWise, ISD, MTy);
800 return LT.first * AVX1CostTblNoPairWise[Idx].Cost;
803 if (ST->hasSSE42()) {
804 int Idx = CostTableLookup(SSE42CostTblNoPairWise, ISD, MTy);
806 return LT.first * SSE42CostTblNoPairWise[Idx].Cost;
810 return TargetTransformInfo::getReductionCost(Opcode, ValTy, IsPairwise);
813 /// \brief Calculate the cost of materializing a 64-bit value. This helper
814 /// method might only calculate a fraction of a larger immediate. Therefore it
815 /// is valid to return a cost of ZERO.
816 unsigned X86TTI::getIntImmCost(int64_t Val) const {
823 return 2 * TCC_Basic;
826 unsigned X86TTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
827 assert(Ty->isIntegerTy());
829 unsigned BitSize = Ty->getPrimitiveSizeInBits();
833 // Never hoist constants larger than 128bit, because this might lead to
834 // incorrect code generation or assertions in codegen.
835 // Fixme: Create a cost model for types larger than i128 once the codegen
836 // issues have been fixed.
843 // Sign-extend all constants to a multiple of 64-bit.
846 ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
848 // Split the constant into 64-bit chunks and calculate the cost for each
851 for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
852 APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
853 int64_t Val = Tmp.getSExtValue();
854 Cost += getIntImmCost(Val);
856 // We need at least one instruction to materialze the constant.
857 return std::max(1U, Cost);
860 unsigned X86TTI::getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
862 assert(Ty->isIntegerTy());
864 unsigned BitSize = Ty->getPrimitiveSizeInBits();
865 // There is no cost model for constants with a bit size of 0. Return TCC_Free
866 // here, so that constant hoisting will ignore this constant.
870 unsigned ImmIdx = ~0U;
872 default: return TCC_Free;
873 case Instruction::GetElementPtr:
874 // Always hoist the base address of a GetElementPtr. This prevents the
875 // creation of new constants for every base constant that gets constant
876 // folded with the offset.
878 return 2 * TCC_Basic;
880 case Instruction::Store:
883 case Instruction::Add:
884 case Instruction::Sub:
885 case Instruction::Mul:
886 case Instruction::UDiv:
887 case Instruction::SDiv:
888 case Instruction::URem:
889 case Instruction::SRem:
890 case Instruction::And:
891 case Instruction::Or:
892 case Instruction::Xor:
893 case Instruction::ICmp:
896 // Always return TCC_Free for the shift value of a shift instruction.
897 case Instruction::Shl:
898 case Instruction::LShr:
899 case Instruction::AShr:
903 case Instruction::Trunc:
904 case Instruction::ZExt:
905 case Instruction::SExt:
906 case Instruction::IntToPtr:
907 case Instruction::PtrToInt:
908 case Instruction::BitCast:
909 case Instruction::PHI:
910 case Instruction::Call:
911 case Instruction::Select:
912 case Instruction::Ret:
913 case Instruction::Load:
918 unsigned NumConstants = (BitSize + 63) / 64;
919 unsigned Cost = X86TTI::getIntImmCost(Imm, Ty);
920 return (Cost <= NumConstants * TCC_Basic) ? TCC_Free : Cost;
923 return X86TTI::getIntImmCost(Imm, Ty);
926 unsigned X86TTI::getIntImmCost(Intrinsic::ID IID, unsigned Idx,
927 const APInt &Imm, Type *Ty) const {
928 assert(Ty->isIntegerTy());
930 unsigned BitSize = Ty->getPrimitiveSizeInBits();
931 // There is no cost model for constants with a bit size of 0. Return TCC_Free
932 // here, so that constant hoisting will ignore this constant.
937 default: return TCC_Free;
938 case Intrinsic::sadd_with_overflow:
939 case Intrinsic::uadd_with_overflow:
940 case Intrinsic::ssub_with_overflow:
941 case Intrinsic::usub_with_overflow:
942 case Intrinsic::smul_with_overflow:
943 case Intrinsic::umul_with_overflow:
944 if ((Idx == 1) && Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
947 case Intrinsic::experimental_stackmap:
948 if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
951 case Intrinsic::experimental_patchpoint_void:
952 case Intrinsic::experimental_patchpoint_i64:
953 if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
957 return X86TTI::getIntImmCost(Imm, Ty);